Repository navigation
Replies: 4 comments
|
Thanks for this, @j0nsh1n. Getting both Model M and RR running under Proton this soon after the source went up is great work, and the reports are detailed enough to act on directly. Where each item stands:
Each fix will come as its own pull request, linked to its issue and crediting you. If you'd rather send any of them yourself, pull requests are welcome; CONTRIBUTING.md has the details. I'll also follow the dxil-spirv change. Once it is in a Proton release, the Linux steps from this post can go into the docs. |
|
Update, @j0nsh1n: the fixes from your reports are on
None of this was run under Wine or Proton, or in a game yet; everything above was measured on Windows with test programs on an RX 9070 XT. If you can rebuild from These fixes are in the source only so far; a 0.6.1 build follows after an in-game test. |
|
Thanks for turning this thread into issues so quickly, and for the credit in the 0.6.1 notes. Here are the follow-ups, each where it belongs:
One vkd3d-proton build (vkd3d-proton's main |
|
Thank you, @j0nsh1n, this is exactly the kind of follow-up that moves things. Yes to the pull requests:
A few practical points for the pull requests. CONTRIBUTING.md has the details: the clean-room statement in your first pull request and a On timing: 0.6.1 is being built now as a bug-fix release with the leak and crash fixes, so the Linux launcher changes will land after it, in the next one. |
Uh oh!
There was an error while loading. Please reload this page.
Model M and Ray Reconstruction both run in Cyberpunk 2077 under GE-Proton 11-11 on an RX 9070 XT (RADV, Mesa 27.0.0-devel, GE-Proton's vkd3d-proton 3.1.0). Getting there needed five fixes, in three projects. This post lists them, how each was checked, and what is still open.
What blocked each part
globallycoherent, which Wine'sd3dcompiler_47(vkd3d-shader) rejectsconvert.cppFailed DXIL opcode 79): 32 of 41 kernelsstd::mutexlock with msvcp140 older than 14.40How the fixes were checked
clean/runtime/interop/tests/run-tests.shunder GE-Proton 11-11 fails 31 of 31 before the spin-shader fix and passes all 31 after it.rr_test_host tagloopjob grows video memory by 2,635 MiB in 200 frames before the fix and 37 MiB after.executor dx12on the game's device. RR denoises and upscales 1706x960 to 2560x1440 at about 7.5 ms (accurate) or 6.1 ms (fast) of GPU time per frame. A 6.5-minute session ran 15,203 RR evaluates with flat memory.Running it today
Until the dxil-spirv change is released in vkd3d-proton and then Proton, you need a self-built vkd3d-proton:
dxgi. Wine's own DXGI reports the card as "ATI Radeon HD 5600 Series", and the launcher refuses it.0bb64c5) with ags: Support WMMA operands packed by the shader HansKristian-Work/dxil-spirv#362. Put the resultingd3d12.dllandd3d12core.dllnext to the game exe. Proton re-copies its own DLLs into the prefix on every launch, so a prefix copy does not last.winetricks vcrun2022, or protontricks), until WrongGPU-RR.asi crashes with msvcp140.dll older than 14.40 (constexpr std::mutex, _Mtx_lock null read) #9 is fixed.wronggpu_runtime.dllfails to compile the spin shader, so Model M falls back to FSR 2.1.2. It also leaks video memory as soon as RR is reported available. A runtime built from source with the changes described in those issues works.Still open
resolve.fixup, takes about 130 seconds to compile through RADV on a cold cache. Until all RR kernels are ready, frames pass through undenoised, which looks like heavy speckle. Later starts use vkd3d-proton's pipeline cache.native_driver_probemultiplies all-ones by all-twos and adds a per-lane value. It passes with any element layout, so it would not catch a layout difference. A probe with distinct values per lane would.NativeD3D12Device::arena_buffer. It is stable, not a leak. Is that size expected?All reactions